command line interface about scikit-learn

[GaelVaroquaux]

I am really not too enthusiastic about such a proposal for the following reasons:

1. Will create a large volume of boiler plate code, which will be a maintenance burden
2. Hard to test
3. Will get us users that do not want to learn Python, and thus give us questions that we cannot answer, non informative bug reports, and force us to write much more documentation

You may think that I am cynical, but I am trying to think in the long run and to make sure the that the project doesn't implode under its own weight.

[mblondel]

I'm not quite sure what kind of boilerplate you're thinking of. I expected the command line program to be standalone and quite small, actually.

Also, since the command would use pickle for persistence, this would mean that people can apply a few pre-processings (feature extraction, PCA, ...), get their pickle object and work from there, in Python.

So I guess the only of your arguments I really agree with is 2.

This feature is not a must for me so if people don't like it too much, no problem!

[GaelVaroquaux]

Maybe I am wrong, but I expect the boiler plate code to come from impedance matching Python with a command line.

For point 3, I guess that my answer to your answer is that users that know Python and need to call the scikit via a command line (eg to work in a multi-language environment) can cook up the functionality they need very quickly.

I am not in favor at all of this feature as I think that it is extending a bit outside of the scope of the scikit, but as always, I can be convinced, if I see that enough developers feel strongly about this and would maintain it.

[ogrisel]

I think this a really important feature for day-to-day practitioners who are not necessarily developers but more data annalists who want to quickly evaluate the output of algos implemented in the scikit on their own data without having to write boilerplate code themselves.

It will be even more important once we implement online API to be able to naturally handle infinite byte streams in a Unix pipe.

[ogrisel]

Having the ability to quickly wrap algorithms and predictive models as Unix CLI tools that read stdin and write to stdout would also make it trivial to use the scikit in a Hadoop Streaming environment (or using Apache Pig with the STREAMING command as well).

[ogrisel]

As for the priority I agree this is not a high priority task: we need to work on the online part first to make this really useful in practice IMHO.

[larsmans]

I think this is something that should be pioneered in a separate package. I feel like closing this issue as I don't see it happening any time soon (and the issue tracker is filling up with "we should implement such and such" as well as PRs).

[amueller]

+1

[amueller]

Thanks for doing the clean-up round ;)

[dan-blanchard]

I believe we've created that separate package you're looking for at ETS, and we just publicly released it on Friday! We called in SciKit-Learn Laboratory (SKLL). You can install from pip with just pip install skll.

Documentation: http://scikit-learn-laboratory.readthedocs.org
Github project: https://github.com/EducationalTestingService/skll

It lets you easily run experiments using a variety of classifiers/regressors when you have pre-generated feature files. We hope other people find it useful, and feedback is always welcome. We use it a lot internally.